Detaching Resources: Detached Search Service

In the previous post, we have reviewed the basics about virtualization, learned about Docker and set up a container running RabbitMQ. In this post, we will create a web app with a search service. The interesting part is that the former will be completely decoupled from the latter.

Regardless of being very contrived, the following example gives us a pretty good bird’s-eye view and the basics for building web apps towards a service-oriented architecture.

A bit more Docker

First we will need to build a container able to run our ruby services. Thanks to docker’s abilities of reusing images, we can create a new image to suit our needs pretty easily by inheriting from another image.

I used this image as the base for my experiment, and added on top of it essential packages like curl, monit and openssh-server. You can my new image here.

Given our goal is to create a fully distributed system, we will start three docker containers, each for the following services: the message queue, the web app and the search engine.

Every time you start a container, docker assigns it a new IP address. This means we have to get the IP address—via docker inspect—of the queue container, to set it on the other containers.

To avoid this intense manual labor, we can use pipework to set up a private network. We will define static IP addresses for the containers and share them using environment variables.

# spin up containers
$ QUEUE=$(docker run -d wicz/rabbitmq)
$ WEB=$(docker run -e QUEUE_HOST= -d wicz/ruby)
$ SEARCH=$(docker run -e QUEUE_HOST= -d wicz/ruby)
# set IP addresses for containers
$ pipework br1 $QUEUE
$ pipework br1 $WEB
$ pipework br1 $SEARCH
# enable host to access the private network
$ ifconfig br1

Docker 0.7 has a new feature to “link” containers together. With this, we can skip the pipework set up, because docker will populate the environments variables with the IP addresses it automatically assigns and with the exposed ports.

Testing the services

With the ruby containers and the message queue running, we can set up the web and search services. We will clone the repository in both containers and start the respective services.

First we need to clone the repository in both web and search containers:

$ ssh root@10.11.12.{2,3} git clone /opt/dr

For the search container we need to start the consumer daemon:

search$ cd /opt/dr/v2/search
search$ bundle exec ruby bin/search_consumer.rb

And for the web container we need to start the consumer and the web server:

web$ cd /opt/dr/v2/web
web$ bundle exec ruby bin/web_consumer.rb
web$ bundle exec ruby web.rb -o &

Now point your browser to the web server and you should see a simple page with a search form. Whatever search you do will redirect to another page with “Searching…” written on it. Wait a few seconds, refresh this page and you should see “It works!”.

Since I am running the docker container inside a Vagrant VM, I forwarded a local port to the web server in the web container to be able to use my browser in the OS X.

$ vagrant ssh -- -L 8001:
# open http://localhost:8001

Behind the scenes

The example is naive, not to mention it always returns the same result for every search. But the significance in it, is to note how the results are generated and the services exchange messages, and not what they are.

When you hit the “Submit” button, the web app creates a new Search object and invokes the SearchService to send it to a queue named dr.searches. Then the client is redirect to a page with the search results. Since it has not been processed yet, it shows the “Searching…” placeholder.

On the other container, the search daemon consumes messages from dr.searches queue. It unserializes the message, changes its content and send it back to another queue named dr.searches.results.

The SearchConsumer knows nothing about the web app, the database SearchRepository or even the Search object. All it knows is how to handle JSON data. Its perfect because we have created a completely decoupled solution. As long as we do not change the JSON specification, we can change whatever we want in the web app and the search engine will continue working. We can also create other apps that use the same search engine.

Back in the web container, the SearchResultsConsumer receives the search results from the dr.searches.results queue. It unserializes the JSON and updates the results in the database.


We have successfully set up our detached services solution. With a few amendments in the business logic, I think it could be used in production environments. If you need a similar solution and do not have the time or energy to improve this one, I suggest giving Hutch from GoCardless a try.

Given the message queue is fully working, we could easily set up another services like long running jobs, e.g. image manipulation or reports generation, by just creating new consumers.

Try this examples yourself. I would love to hear your thoughts on the subject. Comments are open!